Very Fast Image Retrieval Based on JPEG Huffman Tables

In this paper, we present a very fast method for performing content-based image retrieval of JPEG compressed images. Our method works directly in the compressed domain of JPEG but in contrast to previous techniques does not require to undo the entropy coding stages of the compression. The reason for this is that, as we show, image adapted Huffman tables can be directly employed as image descriptors and image similarity defined as similarity between the corresponding tables. To do this, we extract both DC and AC Huffman tables and compare the lengths of assigned prefix codes, which give an indication of the frequencies of the related DC and AC values, and use this to compare images. Since the Huffman tables reside in the header of JPEG images, our approach is extremely fast, as not only does it not require any kind of decompression it also needs reading in only a fraction of the file. We evaluate our method on benchmark databases of varying sizes up to in excess of 1 million images, and show that our approach achieves retrieval performance similar to other techniques, while providing a speedup more than 30-fold compared to JPEG compressed domain algorithms and more than 150-fold compared to common pixel domain techniques for online image retrieval.